Alexandria Engineering Journal (2011) 50, 137144

Alexandria University

Alexandria Engineering Journal

www.elsevier.com/locate/aej
www.sciencedirect.com

ORIGINAL ARTICLE

A methodology for Electric Power Load Forecasting

Eisa Almeshaiei a, Hassan Soltan

b,*

Production Engineering Department, College of Technological Studies, Public Authority for Applied Education and Training,
Sheiwck, Kuwait
b
Production Engineering and Mechanical Design Department, Faculty of Engineering, Mansoura University,
Mansoura 35526, Egypt
Received 21 January 2010; accepted 26 January 2011
Available online 27 July 2011

KEYWORDS
Electric Power Load Forecasting;
Time series;
Pattern segmentation/
decomposition

Abstract Electricity demand forecasting is a central and integral process for planning periodical
operations and facility expansion in the electricity sector. Demand pattern is almost very complex
due to the deregulation of energy markets. Therefore, nding an appropriate forecasting model for
a specic electricity network is not an easy task. Although many forecasting methods were developed, none can be generalized for all demand patterns. Therefore, this paper presents a pragmatic
methodology that can be used as a guide to construct Electric Power Load Forecasting models. This
methodology is mainly based on decomposition and segmentation of the load time series. Several
statistical analyses are involved to study the load features and forecasting precision such as moving
average and probability plots of load noise. Real daily load data from Kuwaiti electric network are
used as a case study. Some results are reported to guide forecasting future needs of this network.
2011 Faculty of Engineering, Alexandria University. Production and hosting by Elsevier B.V.
All rights reserved.

1. Introduction

* Corresponding author.
E-mail addresses: eamboabood@yahoo.com (E. Almeshaiei), hasoltan@
hotmail.com (H. Soltan).
1110-0168 2011 Faculty of Engineering, Alexandria University.
Production and hosting by Elsevier B.V. All rights reserved.
Peer review under responsibility of Faculty of Engineering, Alexandria
University.
doi:10.1016/j.aej.2011.01.015

Production and hosting by Elsevier

Electricity as a product has very different characteristics

compared to a material product. For instance, electricity
energy cannot be stored as it should be generated as soon as
it is demanded. Any commercial electric power company has
several strategic objectives. One of these objectives is to provide end users (market demands) with safe and stable electricity. Therefore, Electric Power Load Forecasting (EPLF) is a
vital process in the planning of electricity industry and the
operation of electric power systems. Accurate forecasts lead
to substantial savings in operating and maintenance costs, increased reliability of power supply and delivery system, and
correct decisions for future development. Electricity demand
is assessed by accumulating the consumption periodically; it

138
is almost considered for hourly, daily, weekly, monthly, and
yearly periods.
The EPLF is classied in terms of the planning horizons
duration: up to 1 day/week ahead for short-term, 1 day/week
to 1 year ahead for medium-term, and more than 1 year ahead
for long-term [2,19]. Short-term forecasts are used to schedule
the generation and transmission of electricity. Medium-term
forecasts are used to schedule the fuel purchases. Long-term
forecasts are used to develop the power supply and delivery
system (generation units, transmission system, and distribution
system). Also, refer to Weron [29] and Pedregal and Trapero
[19].
The electricity demand pattern is necessarily affected by
several factors including time, social, economical, and environmental factors by which the pattern will form various complex
variations [1,10]. Social (such as behavior) and environmental
factors are big sources of randomness (noise) found on the
load pattern. Diversity and complexity in demand pattern have
been leading to developing complicated EPLF methods. The
literature is enriched with EPLF methods having many attempts to nd the best estimation of load forecasting. The
major methods include time series such as exponential smoothing, ARMA, BoxJenkins ARIMA, regression, and transfer
function (dynamic regression); expert systems; neural networks; Fuzzy logic; and support vector machine [4,20,21,
27,29,3,10,18,25]. Recently, ant colony optimization is applied
to EPLF [15]. Furthermore, there are many review papers on
EPLF methods such as Alfares and Nazeeruddin [2], Singh
and Singh [22], Metaxiotis et al. [13], Taylor et al. [24], and
Hahn et al. [9].
The ARIMA models and their versions have achieved a
considerable success for EPLF (see [7,29], and [16]. In general,
ARIMA models can be used when the time series is stationary
without missing data [8]. They can be further hybridized with
articial intelligence techniques [26].
However, the complexity of demand pattern depends on its
base period; it changes from fairly smooth curve (annually
based) to most noisy and cyclic complex curve (hourly based)
since the effect of environmental factors increases. Combined
forecasting was also introduced based on a certain linear combination of various results from different forecasting methods
[5,30,31,28]. Hybrid methods were also developed for EPLF
such as Niu et al. [14,15]. Also, see Weron [29] for an extensive
review of EPLF and its applications.
In this era the electric power consumption is growing fast
and may be more randomly because of the increasing effect
of environmental and human behavior. Therefore, the electricity demand pattern becomes more complex and unrecognized.
For instance, the people all over the world are using increased
number and variety of electric appliances most of them are
environmentally related, that increases the cyclic variation
and noise on the demand pattern. Though there are many forecasting methods, no single one can be generalized to perform
enough for all cases, especially when many factors are considered. Thus, to get a proper forecast, it is not just adopting a
famous method. In other words, an ideal method for a case
may perform poorly for another one. Therefore, the research
must be directed to specially assigned methods. In other words,
each electric power plant in any country needs to follow its
own EPLF method. (That is the direction of this paper.) For
that purpose, the general methods can be also adopted but
with efcient and effective modications that suit the case;

E. Almeshaiei, H. Soltan
otherwise the results will be misleading. For instance, Ohtsuka
et al. [16] have considered an ARMA process as a time series
structure, for their special model, because it is suitable for
the Japanese load data as they stated.
The aim of this paper is to demonstrate a pragmatic EPLF
methodology for analyzing the electric load pattern and predicting the future load demand for short, medium, and/or long
terms. This methodology can integrate different forecasting
models. The rest of the paper is organized as follows. Section
2 is a preliminary for the proposed methodology which is presented in Section 3. Section 4 applies the methodology to a typical power load pattern. Concluding remarks are contained in
Section 5.
2. Preliminaries
Time series forecasting methods and their modications (e.g.
combinations and hybridizations) are the most popular methods for forecasting in general. De Gooijer and Hyndman [7]
have introduced a magnicent general review for time series
forecasting. Electricity load pattern is principally a time series.
Time series analysis is normally the main method used in the
eld of EPLF. As any time series, the load pattern can be modeled as a function of different factors [17,29,19]. Such function
should comprise several components; each component can
meet one or more factors (time, social, economical, or environmental factor). It has been suggested in the literature that time
series of electricity load may contain a trend component [3].
This could take the form of either a deterministic or a stochastic trend. Such functional relationships can assume different
forms such as additive, multiplicative, or mixed forms. An
additive function and a multiplicative function for electricity
load were presented by Chen et al. [6] as
Lt Ln t Lw t Ls t Lr t

and
Lt Ln tFw tFs tFr t

where L(t) is the total load at time t; Ln(t) is the normal or

trend component, which is a set of standardized load shapes;
Lw(t) is the weather sensitive component; Ls(t) is a special
event component that create a substantial deviation from the
usual load pattern; and Lr(t) is a completely random term
(the noise). The Fs are positive correction factors justied in
the same way.Thus, to nd a proper forecast of future electricity load, the behavior of the given pattern must be studied very
well. For instance, when the pattern is a multi-component time
series, the impact of each component on the time series must be
identied especially the noise component. It is obvious that
each component itself is a time series. This necessitates decomposition of the time series in some way to identify and extract
the partial loads (components) that superimposed to the main
load time series and to each other.
The classical decomposition of time series is often based on
four components: trend, cyclic, seasonal, and random components. The seasonal component is viewed as a periodic component with a xed cycling period corresponding to the
individual seasons. In practice, it is convenient to combine
the trend and the cyclical components into a trend-cycle component to reduce to three components time series. Decomposition analysis is often used for detrending and deseasonalizing
the time series. Review Palit and Popovic [17]. The same was

A methodology for Electric Power Load Forecasting

139

frequently followed to analyze the electricity load [29]. Since

the load time series may contain many more components,
the decomposition task becomes very hard. However, for complex load pattern, the decomposition process may fail to nd
proper forecasts.
The global models for EPLF often fail to forecast near
accurate loads due to the nature of time series of electricity
load. Strictly, the load pattern is very complicated non-stationary time series. Therefore, the principle of partitioning the load
time series into enough homogenous regions was introduced in
the literature aiming at improving the accuracy of load forecasting [23,11,12]. After partitioning, each region can be analyzed separately and the relationship between those regions
should be identied. In other words, a different model can

Input Timely Loads

be assigned to each region with monitoring the behavior of

model parameters relative to other models.
3. The proposed EPLF methodology
This section presents the proposed EPLF methodology for
forecasting future load without the restriction to the term
length (short, medium, or long). This methodology is mainly
based on the principles of time series segmentation and decomposition. Some additional statistical analysis will be followed to
aid the decision making based on the adopted forecasts such as
probability plots. The methodology can be presented through
the owchart shown in Fig. 1. The main phases of the methodology can be summarized as follows.

Plot Load Time Series

Investigate Load Features

Begin

Construct Contours Relating Similar Points

Choose Moving Average Base

Choose Smallest Moving Average Order k

Compute Moving Average

Plot Moving Average

Stop
Compute Moving Average Error

Analyze Probability Distribution of Error

k= k + 1

No

Smooth Moving
Average
Yes

Adopt Moving
No
Average Forecast
or Use ARIMA

Dividable
Time Series

Study Possibility and Way of

Combining Current Forecasts
Find the Forecast by Region,
Contour, and Proposed Points
Find Points Proposed Related
to the Required Forecast
Find Contour Nearest to
the Required Forecast
Find Region Similarity of
the Required Forecast

Yes
Stop

Divide Time Series into Homogeneous

Regions Based on Smooth Moving Average

For Each Group: Search

Fitting Fits Parameters

For Each Region: Repeat Searching Smooth

Moving Average

Group Similar Regions

For Each Region: Identify Nearest Fit Type

Based on Smooth Moving Average

For Each Region: Estimate

Fit Based on Original Data

Figure 1

Flowchart for the proposed EPLF methodology.

140

E. Almeshaiei, H. Soltan

Daily load in Kuwaiti network during 20062008.

Figure 2

Figure 3

Figure 4

Daily meshed plot for load during 20062008.

The 7-day moving average for load during 20062008.

3.1. Primal visual and descriptive statistical analysis

3.3. Load pattern decomposition

This gives a primal conclusion about the behavior of the features of load time series.

This is based on only two components; the noise superimposed on the main load component. Moving average, exponential smoothing, or ARMA/ARIMA models can be used
for that purpose. Here, moving average is proposed. The
main component is expected to be smooth. Therefore, moving average order is experimented. The smoothed moving
average leaves a noise component with smallest mean value

3.2. Contour construction

This locates similar points along the time series around a single
trend line, if that is available.

A methodology for Electric Power Load Forecasting

Figure 5

141

Noise based on 7-day moving average for load during 20062008.

Figure 6

The 30-day moving average for load during 20062008.

Figure 7

Trend of the rst region of year 2006.

(near to zero) and standard deviation greater than zero. The

decomposition process is carried out for original time series
and each homogeneous region. It is proposed here that the
moving average base can be a suitable measure of central tendency. The order of moving average should be chosen
carefully.
3.4. Load pattern segmentation
This is based on the decomposition of original time series into
two components to identify segments (regions) of time series.
Each region is identied by a set of points demonstrates some
homogeneity (approximate a trend). This can be done using
analysis of variance or other methods such as self organizing
maps [11,12]. Furthermore, this process can be done visually

or by using simpler statistics. Notice that, in rare cases, the original time series itself may be found homogeneous.
3.5. Future load forecasting
This can be achieved aided by one or more of three bases region similarity, contour, and proposed related points and
their combined forecast.
Curve tting is conducted to each region. The parameters
of resulting trends are studied to nd if there is a correlation
between the parameters of some regions. The regions that have
correlation between their parameters are considered similar
and assigned into a group. Hence, we can forecast the growth
of these parameters. The effect of each region, contour, and
proposed related point on the required forecast can be

142

E. Almeshaiei, H. Soltan

Figure 8

Trend of the second region of year 2006.

0.999
Mean
StDev
N
AD
P-Value

0.99
0.95

1.623177E-10
7230
285
0.406
0.349

Probability

0.8
0.6
0.4
0.2
0.1

0.01
0.001
-30000

-20000

-10000

10000

20000

30000

Noise
Figure 9

Probability plot for noise of rst region of year 2006.

weighted. Thus, the process of estimating a forecast can be

aided with a weighting decision support system. Other similar
systems such as ranking systems can also be used.
The performance of any forecasting methodology can be
evaluated using several proposed statistical metrics (see [29]
and [17]. Most common and reasonable metric is the mean
absolute percentage error that is

n 
1X
jyt  y^tj
MAPE
3
n i1
yt
where y(t) and y^t are the estimated load and actual load at
time t; and n is the number of data used. Less MAPE means
more precise estimates. However, the performance metric suitability depends on the features of the load time series.
4. Analysis of electric load case
The proposed methodology is basically a practical guide to design specic EPLF models for forecasting future electricity
consumption. As cited before, an accurate forecast can not
be obtained blindly through global methods. Thus, this methodology is case dependent. For demonstration, the case of daily loads data of Kuwaiti electric network during 20062008 is

analyzed. These data are obtained from the actual records in

MWh. Original time series of the load is depicted in Fig. 2. Excel 2007 and MINITAB 14 will be used as aids for analysis. It
is obvious that the load shows repetitive cyclic features for
nearly equal lengths. Three main types of contours can be
drawn. The upper, middle, and lower contours are approximately increasing straight lines. These contours can be used
to forecast some special days loads. Also, other contours
can be drawn.
Fig. 3 illustrates the meshed plot for the load during the
three years. This gives sight about the growth of daily loads,
which increases in most of days. This plot can be used for load
weighting. The simple moving average is used for data ltering. It is found that the 7-day moving average (Fig. 4) is the
nearest smooth one with mean error about 30.55 MW and
MAPE of 0.0384. This error represents the noise (Fig. 5) based
on moving average. Test of hypothesis proved that the current
noise can not be said normally distributed with having the
apparent odd values. This moving average seems sufcient to
discover the regional homogeneity of the time series. Further
smoothing will not yield the actual noise because of starting
to distort the main load. For instance, 30-day moving average
(Fig. 6) gives more information about homogeneity with mean

A methodology for Electric Power Load Forecasting

143

225000
S
R-Sq
R-Sq(adj)

200000

7268.21
97.2%
97.2%

175000

MW

150000

125000

100000

75000

50000
1

30

60

90

120

150

180

210

240

270

300

Day

Figure 10

Fitted polynomial plot with condence 95% for rst region of year 2006.

error of 174.47 MW, while this error cant be considered as

noise. As shown later, the segmentation process will yield normally distributed noise with mean about zero.
Referring to Figs. 4 and 6, the load time series of Kuwaiti
network for the given period can be divided into six regions,
each year into two regions. The rst region of each year is
around 285 days (about 9.5 months) horizon. All regions can
be tted to third order polynomial as shown for instance in
Figs. 7 and 8 of year 2006. The rst regions (rst group) have
a specic similarity and the second regions (second group)
have another similarity. In other words, for each group, each
polynomial parameter demonstrates a trend. Probability plots,
as that shown in Fig. 9, are constructed to test the hypothesis
of the noise probability distribution based on the estimated
trend polynomials. The noise along the time series can be assumed normally distributed with mean about zero. Therefore,
a forecast estimate can be obtained with a specic condence
level, as shown in Fig. 10. Fortunately, in this case study, each
region can be used to forecast a daily future load by applying
slight modication to the trend parameters.
5. Concluding remarks
Electricity demand forecasting represents the main task in the
planning of electricity production because it determines the required resources to operate the electricity plants such as daily
consumption of fuels. Furthermore, it is the corner stone of
planning for electric plants and networks. The literature reports that the electric load pattern is very complex. It is therefore necessary to develop new methods for EPLF to reduce the
uncertainty of the predictions. This paper has reported that
every electric network and plant needs to its special forecasting
method because each country is indifferent in the factors that
affect the electricity demand. In the developing countries the
electric demand grows with dynamic and high growth rate.
The methodology of this paper is developed as a guide for
constructing specialized EPLF models. It proposes three bases
region similarity, contour, and proposed related points and
their combined forecasts. Forecasting analysis is conducted to

the daily load time series of Kuwaiti electric network during

three years. The analysis proved that the proposed methodology accommodates this case study. The segmentation process
results in homogeneous regions for which polynomial trends
have been identied. Furthermore, it is found that the superimposed noise can be considered normally distributed, which enable constructing condence intervals for future forecasts.
References
[1] M.R.G. Al-Shakarchi, M.M. Ghulaim, Short-term load
forecasting for baghdad electricity region, Electric Machines
and Power Systems 28 (2000) 355371.
[2] H.K. Alfares, M. Nazeeruddin, Electric load forecasting:
literature survey and classication of methods, International
Journal of Systems Science 33 (2002) 2334.
[3] L.F. Amaral, R.C. Souza, M. Stevenson, A smooth transition
periodic autoregressive (STPAR) model for short-term load
forecasting, International Journal of Forecasting 24 (2008) 603
615.
[4] C. Alvarez, S. An, Stochastic load modeling for electric energy
distribution applications, Top 2 (1994) 151166.
[5] G.J. Chen, K.K. Li, T.S. Chung, H.B. Sun, G.Q. Tang,
Application of an innovative combined forecasting method in
power system load forecasting, Electric Power Systems Research
59 (2001) 131137.
[6] H. Chen, C.A. Canizares, A. Singh. ANN-based short-term load
forecasting in electricity markets. in: Proceedings of the IEEE
Power Engineering Society Transmission and Distribution
Conference 2, 2001, p. 411415.
[7] J.G. De Gooijer, R.J. Hyndman, 25 Years of time series
forecasting, International Journal of Forecasting 22 (2006)
443473.
[8] V.S
. Ediger, S. Akar, ARIMA forecasting of primary energy
demand by fuel in Turkey, Energy Policy 35 (2007) 17011708.
[9] H. Hahn, S. Meyer-Nieberg, S. Pickl, Electric load forecasting
methods: tools for decision making, European Journal of
Operational Research 199 (2009) 902907.
[10] H.R.S. Keyno, F. Ghaderi, A. Azade, J. Razmi, Forecasting
electricity consumption by clustering data in order to decrease
the periodic variables effects and by simplifying the pattern,
Energy Conversion and Management 50 (2009) 829836.

144
[11] M. Mart n-Merino, J. Roman, A new SOM algorithm for
electricity load forecasting, in: I. King et al. (Eds.), ICONIP
2006, Part I, LNCS 4232, Springer-Verlag, 2006, pp. 9951003.
[12] M. Mart n-Merino, J. Roman, Electricity load forecasting using
self organizing maps, in: S. Kollias et al. (Eds.), ICANN 2006,
Part II, LNCS 4132, Springer-Verlag, 2006, pp. 709716.
[13] K. Metaxiotis, A. Kagiannas, D. Askounis, J. Psarras, Articial
intelligence in short term electric load forecasting: a state-of-theart survey for the researcher, Energy Conversion and
Management 44 (2003) 15251534.
[14] D.-X. Niu, Q. Wang, J.-C. Li, Short term load forecasting model
based on support vector machine, in: D.S. Yeung et al. (Eds.),
ICMLC 2005, LNAI 3930, Springer-Verlag, 2006, pp. 880888.
[15] D. Niu, Y. Wang, D.D. Wu, Power load forecasting using
support vector machine and ant colony optimization, Expert
Systems with Applications 37 (2010) 25312539.
[16] Y. Ohtsuka, T. Oga, K. Kakamu, Forecasting electricity
demand in Japan: a bayesian spatial autoregressive ARMA
approach, Computational Statistics and Data Analysis 54 (2010)
27212735.
[17] A.K. Palit, D. Popovic, Computational Intelligence in Time
Series Forecasting: Theory and Engineering Applications,
Springer-Verlag London Ltd., 2005.
[18] S.Sp. Pappas, L. Ekonomou, P. Karampelas, D.C.
Karamousantas, S.K. Katsikas, G.E. Chatzarakis, P.D.
Skadas, Electricity demand load forecasting of the hellenic
power system using an ARMA model, Electric Power Systems
Research 80 (2010) 256264.
[19] D.J. Pedregal, J.R. Trapero, Mid-term hourly electricity
forecasting based on a multi-rate approach, Energy
Conversion and Management 51 (2010) 105111.
[20] I.S. Qamber, E.A. Al-Gallaf, Electric load forecasting model for
the state of Bahrain network, Electric Power Components and
Systems 29 (2001) 259276.
[21] J.V. Ringwood, D. Bofelli, F.T. Murray, Forecasting electricity
demand on short, medium and long time scales using neural

E. Almeshaiei, H. Soltan

[22]

[23]
[24]

[25]

[26]

[27]

[28]

[29]

[30]

[31]

networks, Journal of Intelligent and Robotic Systems 31 (2001)

129147.
D. Singh, S.P. Singh, Self organization and learning methods in
short term electric load forecasting: a review, Electric Power
Components and Systems 30 (2002) 10751089.
H.A. Soltan, On forecasting with blind data, Mansoura
Engineering Journal 25 (2000) M28M34.
J.W. Taylor, L.M. de Menezes, P.E. McSharry, A comparison
of univariate methods for forecasting electricity demand up to
a day ahead, International Journal of Forecasting 22 (2006)
116.
J.W. Taylor, Triple seasonal methods for short-term electricity
demand forecasting, European Journal of Operational Research
204 (2010) 139152.
O. Valenzuela, I. Rojas, F. Rojas, H. Pomares, L.J. Herrera, A.
Guillen, L. Marquez, M. Pasadas, Hybridization of intelligent
techniques and ARIMA models for time series prediction, Fuzzy
Sets and Systems 159 (2008) 821845.
H. Wang, B.-S. Li, X.-Y. Han, D.-L. Wang, H. Jin, Study of
neural networks for electric power load forecasting, in: J. Wang
et al. (Eds.), ISNN 2006, LNCS 3972, Springer-Verlag, 2006, pp.
12771283.
J. Wang, W. Zhu, W. Zhang, D. Sun, A trend xed on rstly and
seasonal adjustment model combined with e-SVR for short-term
forecasting of electricity demand, Energy Policy 37 (2009) 4901
4909.
R. Weron, Modeling and forecasting electricity loads and
prices: a statistical approach, John Wiley & Sons Ltd.,
England, 2006.
Y. Yao, Z. Lian, S. Liu, Z. Hou, Hourly cooling load prediction
by a combined forecasting model based on analytic hierarchy
process, International Journal of Thermal Sciences 43 (2004)
11071118.
P. Zhou, B.W. Ang, K.L. Poh, A trigonometric grey prediction
approach to forecasting electricity demand, Energy 31 (2006)
28392847.